Three-dimensional (3D) cloud retrievals are critical for understanding their impact on climate and other applications such as aviation safety, weather prediction, and remote sensing. However, obtaining high-resolution and accurate vertical representation of clouds remains unsolved due to the limitations imposed by satellite instrumentation, viewing conditions, and the complexity of cloud dynamics. Cloud masks are essential for comprehending various cloud vertical properties, but deriving accurate 3D cloud masks from 2D satellite imagery data is a challenging task. To tackle these challenges, we introduce a physics-informed loss function for training deep learning models that can extend 2D cloud images into 3D cloud masks. The proposed loss, called CloudMask Loss, is composed of two domain knowledge-informed loss terms: one for evaluating cloud position and thickness, and the other for measuring the number of layers. By combining these loss terms, we improve the trainability of the deep learning models for more accurate and meaningful results. We apply the proposed loss function to different neural networks and demonstrate significant improvements in multi-layer cloud mask reconstruction. Utilizing the same neural network architecture, our proposed loss outperforms standard binary crossentropy loss in terms of multi-layer cloud classification accuracy, number of layers accuracy, and thickness mean absolute error (MAE). The proposed loss function can be readily integrated into various neural network architectures, resulting in substantial performance gains in 3D cloud mask generation.